Computer Vision: Teaching Machines to See

Why “seeing” is hard for a computer

You glance at a photo and instantly know it shows a cat. A computer cannot do that directly — it has no eyes and no idea what a cat is. Computer Vision (CV) is the field of AI that lets software make sense of images and video: classify what is in them, find where objects are, and recognise faces.

An image is just a grid of numbers

The first thing to understand: to a computer, a picture is a grid of pixels, and each pixel is just a number for its brightness. A grayscale image uses one number per pixel (0 = black, 255 = white); a colour image uses three numbers per pixel (how much Red, Green and Blue). That is all an image really is.

# A tiny 3x3 grayscale image as numbers (0 = black, 255 = white)
image = [
    [  0,   0,   0],
    [255, 255, 255],
    [  0,   0,   0],
]

# The middle row is bright (255) — that is a white horizontal stripe.
for row in image:
    print(row)

The three core vision tasks

In short: classification gives the picture one label, detection draws labelled boxes around several objects, and recognition matches a face to a specific identity. They get harder going down the list.

How a CNN “sees”

The workhorse of computer vision is the CNN (Convolutional Neural Network) you met earlier. Here is the intuition for how it turns that grid of numbers into a label — built up in layers:

1. It slides small filters (tiny grids of weights) across the image. Each filter lights up when it finds a particular tiny pattern, like an edge or a corner.
2. The next layer combines those edges into simple shapes — curves, circles, textures.
3. Deeper layers combine shapes into parts — an eye, an ear, a wheel.
4. The final layer combines the parts into a whole-object decision — “these two eyes + whiskers + pointy ears → cat.”

So a CNN builds understanding from small to big: pixels → edges → shapes → parts → object. Crucially, it learns which filters to use from thousands of labelled images — nobody programs “a cat has whiskers” by hand.

# What using a trained vision model looks like (concept, not run here)
# A pre-trained model takes the pixel grid and returns label + confidence.
prediction = vision_model.predict(image)

print(prediction)